Machine-Learning Prediction and Reducing Overdoses with EHR Nudges (mPROVEN)

University Of Pittsburgh At Pittsburgh, Pittsburgh PA

Investigators

Linked publications & trials

Abstract

The growing adoption of machine learning (ML)-based risk prediction algorithms in healthcare offers the potential to enhance clinical decision-making and improve patient outcomes. However, with this growth comes a range of ethical challenges to consider during development, testing, and implementation. If left unaddressed, these ethical issues could exacerbate or create disparities in care, lead to potential harm for vulnerable populations, loss of trust among marginalized groups and healthcare providers, or pose legal risks. Our team has conducted pivotal work in developing, externally validating, and implementing machine learning algorithms to identify patients at high risk of opioid overdose. Through this work, we have uncovered and mitigated multiple potential ethical issues. While we have learned key lessons and strategies to improve the implementation of these models, these lessons remain largely as internal institutional knowledge and not documented, organized, or made available publicly to others who seek to implement similar models. The objective of this supplemental application is to address this critical gap by developing and disseminating an educational resource—a ‘playbook’—to guide stakeholders in recognizing and mitigating ethical and implementation challenges associated with ML-based opioid risk prediction tools. Our long-term goal is to promote the equitable and ethical use of artificial intelligence to prevent opioid overdose, reduce health disparities, and improve patient outcomes. Building on our extensive experience, existing literature, and insights gained through stakeholder-engaged discussions, we will develop, refine, and disseminate a playbook that systematically addresses ethical considerations at each phase of developing and implementing ML-based opioid-related risk prediction models. Development of the playbook will follow human-centered design principles in four phases: 1) Draft playbook outline, 2) Iterative refinement of outline, 3) Playbook development and deliverables, and 4) Dissemination. The playbook will be iteratively refined through stakeholder feedback to ensure usability, relevance, and impact and will provide practical guidance to support ethical AI deployment, ensuring responsible and equitable use in healthcare settings. Our broad dissemination strategy will ensure that the playbook is widely available to all. This application aligns well with the key priorities outlined in this Notice of Special Interest (NOT-OD-25-015), particularly in relation to “Research on the ethical issues raised by attempts to maximize translatability and minimize bias in the application of new and emerging technologies.” Additionally, this project will deliver a product that supports NIDA’s priority of leveraging data science and analytics and building bioethics capacity to ensure effective translation, implementation, and dissemination of research findings.